Anthropic details invisible watermarks for Claude text using SynthID

In short: Anthropic says it will add invisible “watermarks” to text written by its Claude chatbot, using a version of Google DeepMind’s open-source SynthID-Text system.

What happened

Anthropic explained how it plans to mark Claude-generated text with an invisible signal that can later be detected. The goal is to meet new European Union rules that require AI-made content to include machine-readable labels.

The company said its approach is based on SynthID-Text, an open-source system from Google DeepMind. “Open-source” means the underlying method is publicly shared so others can study it and build on it.

Anthropic described the basic idea with a simple example. When Claude is writing a sentence, there are often several word choices that mean nearly the same thing, like choosing between “overcast” or “grey.” The system uses many of these small, low-impact choices across a piece of text to leave a hidden pattern, like a faint stamp you cannot see but a scanner can.

Anthropic says the watermark should not change what the text means and should not noticeably affect quality. It also says it will not make Claude more expensive for users.

Anthropic is also adding support for C2PA on images processed by Claude. C2PA is a standard that can attach “content credentials” to media (like an information label) so tools can check where an image came from and whether it was edited.

Why it matters

As AI-written text spreads across the internet, it can be hard to tell what was written by a person and what was produced by a chatbot. Invisible markers could make it easier for platforms, researchers, and regulators to spot AI-generated material, even when it looks normal to readers.

Source: The Verge AI